make find_cuda_exe more lenient

The find_cuda_exe script will reject executables labelled with a CUDA version newer than the version reported by nvidia-smi. This should be changed so that the executable with the nearest available CUDA version is returned.

See e.g. https://docs.nvidia.com/datacenter/tesla/drivers/#cuda-drivers - for the Tesla, nvidia-smi reports CUDA 10.1, but is actually compatible with CUDA 10.2 and 11.x